DXAI: Explaining Classification by Image Decomposition

TL;DR

The paper tackles explainable AI for image classification when discriminative cues are dense or additive by introducing DXAI, a decomposition-based framework that splits an input $x$ into a class-agnostic part $\psi_{Agnostic}$ and a class-distinct part $\psi_{Distinct}$ so that $x=\psi_{Agnostic}+\psi_{Distinct}$. An approximate solution employs multi-branch style-transfer GANs with an $\alpha$-blending mechanism to isolate class-specific features, guided by a pre-trained classifier $C$ and a shared multi-head discriminator that also performs classification. Experiments on diverse datasets demonstrate high-resolution, multi-channel explanations that capture color and texture cues beyond what heatmaps provide, outperforming several baselines on objective fidelity metrics. The approach offers a new lens for XAI, with potential extensions to diffusion-based generators, while trade-offs include training complexity and the lack of a natural pixel-wise importance ranking.

Abstract

We propose a new way to explain and to visualize neural network classification through a decomposition-based explainable AI (DXAI). Instead of providing an explanation heatmap, our method yields a decomposition of the image into class-agnostic and class-distinct parts, with respect to the data and chosen classifier. Following a fundamental signal processing paradigm of analysis and synthesis, the original image is the sum of the decomposed parts. We thus obtain a radically different way of explaining classification. The class-agnostic part ideally is composed of all image features which do not posses class information, where the class-distinct part is its complementary. This new visualization can be more helpful and informative in certain scenarios, especially when the attributes are dense, global and additive in nature, for instance, when colors or textures are essential for class distinction. Code is available at https://github.com/dxai2024/dxai.

Figures (29)

• Figure 1: Our method vs. heatmaps, illustrating three scenarios where heatmaps are less informative: 1) Many details spread across large portions of the image are helpful for accurate classification (top row), heatmaps show only partial relevant information. 2) Distinguishing between types of peppers that differ mainly by color (2nd row). 3) Detecting additive statistical disturbance (a class of clean images and a class of images with noise). Since the contribution is global -- heatmaps face difficulties explaining the reason for classification.
• Figure 2: Three visualizations using a heatmap. Highlighting cars in an aerial image (left). Generated by Grad-CAM selvaraju2017grad.
• Figure 3: An image can often be decomposed into meaningful additive parts, such as in classical structure-texture decomposition (top-left). This example shows how it allows to explain well color-based reasoning.
• Figure 4: Class-distinct component $\psi_{Distinct}$ by our DXAI algorithm and by heatmap manipulation, in which the input image is weighted by the normalized heatmap (see details in \ref{['sec:exp']}). We obtain high resolution, dense multi-channel explanations.
• Figure 5: Training process diagram. The class distinct part (existence/absence of white rectangle) is in the first branch (top, in red), whereas the class agnostic components, which belong to both classes, are generated by subsequent branches. The $\alpha$-blended generation is a major mechanism controlling the training.

Theorems & Definitions (2)

• definition thmcounterdefinition: Class Agnostic
• definition thmcounterdefinition: Class Distinct / Agnostic parts of $x$